HawkEars: A regional, high-performance avian acoustic classifier

Passive acoustic monitoring is rapidly emerging as a dominant approach for studying acoustic wildlife, with neural networks used as an increasingly common and promising approach for extracting detections of particular species from acoustic recordings. Existing options for avian classifiers include small custom models for focal species or large models that attempt to classify the entire global avian community, which suggests a possible tradeoff between classifier performance and species coverage. We argue that building domain-specific classifiers for particular geographic regions provides improved performance in exchange for reduced species coverage and present HawkEars, a regional avian classifier for Canada that includes 314 bird and 13 amphibian species. A major challenge in classifier development is the weak labeling of open access datasets. We developed a novel solution, using embedding-based search to efficiently generate strong labels. We evaluated HawkEars performance for bird species relative to two prominent avian community classifiers: BirdNET, and Perch for two datasets representing two applications: bird community surveys and studies of vocal activity rate. We found HawkEars had substantially higher performance across all metrics, detected on average two more species per recording minute in our community evaluation dataset, and had a recall of nearly twice Perch and four times BirdNET, given a precision of 0.9, for our vocal activity evaluation dataset. We suggest HawkEars provides better classification performance because a smaller species pool allows for more resources allocated per species to training and tuning and reduces the risk of class overlap, and our strong labeling method ensures high-quality training data. While our classifier, HawkEars, is a substantial improvement for practitioners studying acoustic wildlife in Canada and the northern United States, practitioners in other regions can use the HawkEars open-source code to build classifiers for other geographic regions. By continuing to improve deep-learning classification performance, HawkEars has the potential to substantially improve the efficiency and utility of passive acoustic monitoring studies. • HawkEars: A high-performance bird sound classifier for Canada • We used a novel embeddings search to create strongly-labeled training datasets • We incorporated several custom heuristics and a submodel to improve performance • HawkEars outperforms existing classifiers across all metrics at the community level and for more than 80 % of single species

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.